DreamBooth
What is DreamBooth?
DreamBooth is a technique for training an AI image model on a small set of photos of a specific subject so it can generate that subject in new situations, styles, and contexts.
At a glance
- Type of model
- Fine-tuning technique for personalizing existing text-to-image diffusion models
- Developed by
- Google Research
- Key capability
- Training an AI image generation model on three to thirty images of a specific subject to enable generation of that subject in new contexts, poses, and styles
- How it fits in AI workflow
- Used to create custom character models, brand-consistent visual tools, and personalized generators within AI production pipelines; typically applied to Stable Diffusion-based models and workflows
Ready to create?
Direct scenes, design characters, and ship full films
All-in-one AI creative platform with simple, transparent pricing, no speed throttles, and an infinite Canvas for max creativity.
How it compares
DreamBooth produces a full fine-tuned model checkpoint and typically achieves strong, comprehensive personalization of the subject across diverse prompt contexts. LoRA is a more computationally efficient fine-tuning approach that trains a small set of additional weights rather than the full model, requiring less storage and training time while achieving strong but sometimes less comprehensive personalization. In practice, DreamBooth with LoRA combines both approaches, using the DreamBooth training methodology with the LoRA efficiency framework to balance quality against resource requirements.
Pro tip
Image curation for DreamBooth training has a disproportionate impact on output quality. Rather than collecting as many images as possible, prioritize ten to twenty diverse, high-quality images that show the subject from varied angles, in different lighting conditions, and with different backgrounds. Including near-duplicate images, multiple very similar frames, or images with other visually dominant elements teaches the model the wrong patterns. Variety within a small, well-curated set consistently outperforms large sets of redundant images.
Types and variations
- Full DreamBooth fine-tuning updates all or most of the model's weights on the subject dataset, producing comprehensive and flexible personalization but requiring more storage as a full model checkpoint is produced.
- DreamBooth with LoRA integrates the DreamBooth approach with the LoRA efficient fine-tuning framework, reducing storage requirements and training time while maintaining strong personalization results.
- Class-specific DreamBooth training uses prior preservation loss, training the model with additional generic class images to prevent the fine-tuning from degrading the model's general capability while it learns the specific subject.
Ready to make your first scene in Morphic?
Try MorphicCommon use cases
- Training a custom character model from a small set of reference images to generate that character consistently across many different prompts and scenes.
- Creating a brand-specific generation model trained on product imagery, enabling consistent product visualization in any context described in a prompt.
- Personalizing an image generation model with a specific artistic style by training on a curated set of stylistically consistent reference images.
- Building a recurring AI spokesperson or avatar from a photograph set for use across marketing, educational, and communications content.
- Fine-tuning models for domain-specific creative applications where the default base model does not perform well on the specific subjects or styles required.
Ready to create?
Direct scenes, design characters, and ship full films
All-in-one AI creative platform with simple, transparent pricing, no speed throttles, and an infinite Canvas for max creativity.
FAQs
DreamBooth is a fine-tuning technique that trains an existing AI image generation model on a small set of images, typically three to thirty, depicting a specific subject. The trained model can then generate that subject in any context, style, or pose described in a prompt.
DreamBooth can work with as few as three to five images for basic results, but ten to thirty carefully curated, diverse images typically produce more flexible and consistent output. Image quality and variety matter more than volume.
DreamBooth was developed by researchers at Google and described in a paper published in 2022. It has since been widely adopted and adapted by the open-source AI image generation community.
DreamBooth produces a full fine-tuned model checkpoint and typically achieves comprehensive personalization. LoRA trains a smaller set of additional weights that are overlaid on the base model, requiring less storage and training time. DreamBooth with LoRA combines both approaches for a balance of quality and efficiency.
DreamBooth can train models on specific people, characters, products, artistic styles, pets, objects, and any other subject with distinctive visual features that need to be reproducible across diverse generated contexts.
DreamBooth is most commonly applied to Stable Diffusion-based models and their variants, where the open-source model weights can be fine-tuned locally or through cloud training services. It is not applicable to proprietary closed models where the underlying weights are not accessible.
Training time varies with hardware, dataset size, and training parameters, but a standard DreamBooth run typically takes between fifteen minutes and several hours on consumer-grade or cloud GPU hardware. DreamBooth with LoRA generally trains faster than full weight DreamBooth.
Prior preservation loss is a technique used during DreamBooth training where additional generic class images are included alongside the subject images to prevent the fine-tuning from degrading the model's ability to generate the general class of subject. For example, when training on a specific person, generic portrait images are included to prevent the model from forgetting what general portraits look like.
